Claude Certified Architect – Foundations (CCAR-F, also written CCA-F) Practice Test

A coding agent calls tools inside a loop. A developer needs to decide when to terminate the loop. Which is the correct, robust termination signal?

• AScan each assistant message for a completion phrase such as “done” or “task complete” and stop the first time one appears.
• BTreat a ceiling of ten iterations as the loop’s primary stopping condition, whatever state the run is in when it is reached.
• CKeep executing tools and looping while the response’s stop_reason is tool_use; stop when stop_reason is end_turn.
• DEnd the loop as soon as any tool returns an error, since a failed call leaves nothing worth appending.

A team’s codebase-analysis skill returns a long structural report every time it runs. By the third invocation the main session can no longer answer questions about the modules it mapped on the first run, and falls back on generic description of how a service of that kind is usually laid out instead of what this codebase actually contains. Which change addresses the cause?

• AAdd context: fork to the skill’s SKILL.md frontmatter so it runs in an isolated subagent context and returns only its summary to the main session.
• BAdd an instruction to the skill’s SKILL.md body telling it to summarise its findings and keep every report short, so less of it lands in the conversation.
• CRun /compact after each invocation so the accumulated reports are summarised and the session’s token budget is reclaimed before the next run.
• DMove the skill to ~/.claude/skills/ so it becomes a personal variant and its output is scoped to your own session rather than to the project.

You’re designing a tool that queries an external inventory API for an agent. The API sometimes returns “item not found” or rate-limit errors. How should the tool handle these to maximise the agent’s ability to recover?

• AWrite the failure to the server’s own error log with a category and a request ID, so operators can see it while the tool returns nothing to the agent.
• BReturn a structured result describing the error (clear message and error type) as the tool result, so Claude can read it and decide how to proceed.
• CSet the MCP isError flag and pass the provider’s raw exception text through as the message, since the flag already marks the call as failed.
• DRetry the call inside the tool on a fixed interval until it succeeds, so the agent only ever sees a completed lookup.

A review prompt sets out in prose the four fields every finding must carry, and the model still returns findings that drop a field and change shape between runs. Which two changes address that most directly?

• ARestate the four field names a second time immediately before the diff, so the requirement is the last thing the model reads before it generates
• BAdd two or three worked examples that render a complete finding in the exact four-field shape the output is meant to take
• CDescribe each of the four fields at greater length, so the model is given a fuller specification of what each one should contain
• DDefine the finding as a JSON schema on an extraction tool and read the findings out of the tool_use response rather than out of prose
• EMove the four field names into the system prompt, leaving the diff and the instruction to review it in the user turn

Select two answers